Robot lithium battery SOC prediction method based on VDM-CNN-LSTM neural network

Abstract

The invention discloses a robot lithium battery SOC prediction method based on a VDM-CNN-LSTM neural network. Relates to the technical field of neural networks, robots, lithium batteries and the like, and provides an efficient and accurate SOC prediction method for solving the problem that the electric quantity state of a lithium battery needs to be accurately predicted in robot application to prevent over-discharge or damage. The method specifically comprises the following steps: firstly, performing multi-scale decomposition on lithium battery operation data through a VDM, and extracting an intrinsic mode function (IMFs); the spatial relevance of frequency characteristics is captured by using CNN, LSTM modeling time sequence dynamic change is combined, and finally a high-precision SOC prediction value is output through a full connection layer. Experimental results show that the RMSE of the method is reduced to 0.017 under the complex working condition and is improved by 56.4% compared with that of a traditional LSTM, the real-time requirement is met, the method has good robustness and practicability, and high-precision prediction can be kept under the complex working condition.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot lithium battery management, and specifically relates to a method for predicting the state of charge (SOC) of a robot lithium battery based on a VDM-CNN-LSTM neural network. Background Art

[0002] As the core power source of intelligent devices such as robots, the SOC prediction of lithium batteries is an important indicator to ensure the normal operation of the devices. However, due to the non-linear characteristics of lithium batteries and the complexity of the working environment, SOC, as a key parameter to measure the remaining capacity of lithium batteries, is crucial for the efficient operation of robots and other battery-powered devices. In practical applications, accurate prediction of SOC not only helps to avoid overcharging or over-discharging of the battery, thereby extending the battery life, but also can optimize the energy management system of the device and improve the overall performance.

[0003] Currently, some commonly used SOC estimation methods by researchers include the ampere-hour integration method, the model method, the data-driven method, and the coulomb integration method. The ampere-hour integration method has a long-term prediction deviation > 5% due to cumulative errors; the equivalent circuit model has significant parameter drift under temperature fluctuations, and the root mean square error (RMSE) > 0.05; although the traditional LSTM can capture temporal dependencies, it does not consider multi-scale frequency features, and the mean absolute error (MAE) is 0.058. The ampere-hour integration method is relatively simple, but the integration of current will be accompanied by certain errors, which accumulate continuously, and the initial value of SOC needs to be known; the model method predicts SOC based on the electrochemical or equivalent circuit model of the battery, but the model construction is complex, and the parameters are easily affected by temperature, aging, etc. The data-driven method is a machine learning method based on support vector machines (SVM) and random forests. Although the accuracy is improved to a certain extent, the time series information is not fully mined. The coulomb integration method calculates SOC by integrating the current value, but it is easily affected by the initial conditions and cumulative errors, and has a high dependence on the sensor accuracy. Summary of the Invention

[0004] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for predicting the state of charge (SOC) of a robot lithium battery based on a VDM-CNN-LSTM neural network, to achieve high-precision prediction of the SOC of the lithium battery, which specifically includes the following steps:

[0005] Step 1, data collection and preprocessing;

[0006] Step 2, adaptively decompose the preprocessed signal by variational mode decomposition (VDM) to obtain multiple intrinsic mode functions (IMFs), and then optimize the parameters by particle swarm optimization (PSO);

[0007] Step 3: Take the multiple Intrinsic Mode Functions (IMFs) after Variational Mode Decomposition (VMD) as multi-channel inputs. Each IMF is independently input into a Convolutional Neural Network (CNN), and depthwise separable convolutional kernels are used to extract cross-modal coupling features, thereby obtaining the feature sequence output by the CNN;

[0008] Step 4: Input the feature sequence output by the Convolutional Neural Network (CNN) into a Bidirectional Long Short-Term Memory (BiLSTM), and introduce a multi-head self-attention mechanism to strengthen the weights of key time steps;

[0009] Step 5: Map through a fully connected layer and perform inverse normalization to output the SOC prediction value.

[0010] Furthermore, Step 1 is specifically as follows:

[0011] Collect the voltage, current, and temperature data of the lithium battery in real time. To ensure the stability of model training and the scale consistency between different features, time series data is intercepted through a dynamic sliding window, and segmented normalization is performed according to the charge and discharge state and the battery temperature. The normalization formula is as shown in Equation (1):

[0012]

[0013] where phase ∈ {charge, discharge, idle}, μ phase (T) and σ phase (T) are the mean and standard deviation of the corresponding working conditions at temperature T, which are dynamically updated according to the state of health of the battery (SOH).

[0014] Furthermore, Step 2 is specifically as follows:

[0015] Use Variational Mode Decomposition (VMD) to adaptively decompose the preprocessed signal. The number of modes K and the penalty factor α are dynamically adjusted through Particle Swarm Optimization (PSO). The optimization objective is to minimize the spectral overlap and maximize the mutual information between the IMF components and the SOC labels. The decomposition objective function is as shown in Equation (2):

[0016]

[0017] where the number of modes K is adaptively determined according to the state of health of the battery (SOH), as shown in Equation (3):

[0018]

[0019] And, the PSO optimization objective function is as shown in Equation (4):

[0020]

[0021] where MI is the mutual information and λ is the sparsity weight coefficient.

[0022] Further, step 3 is specifically as follows:

[0023] Input multiple Intrinsic Mode Functions (IMFs) into a Convolutional Neural Network (CNN), and use depthwise separable convolution kernels to extract cross-modal frequency features. The convolution operation formula is shown in Equation (5):

[0024]

[0025] where u m is the m-th IMF, M is the number of selected modes, is the cross-channel convolution kernel weight.

[0026] Moreover, the number of parameters of the depthwise separable convolution kernel is shown in Equation (6):

[0027] Params = D k ·M + D k ·D o ·(K ω - 1) (6)

[0028] where D k is the number of input channels, D o is the number of output channels, and K ω is the convolution kernel size.

[0029] Further, step 4 is specifically as follows:

[0030] Use a Bidirectional Long Short-Term Memory (BiLSTM) combined with a multi-head self-attention mechanism to perform time series modeling on the features output by the CNN. The multi-head self-attention mechanism strengthens the weights of key time steps, as shown in Equation (7):

[0031]

[0032] The final hidden state is shown in Equation (8):

[0033]

[0034] where γ is a learnable attention fusion coefficient.

[0035] Further, step 5 is specifically as follows:

[0036] Map to the normalized SOC value through a fully connected layer and perform dynamic inverse normalization based on the real-time temperature, as shown in Equation (9):

[0037]

[0038] where ΔSOC(T) = SOC max (T) - SOC min(T), which is obtained by looking up the battery temperature-capacity curve.

[0039] Compared with the prior art, the advantages of the present invention are as follows:

[0040] 1. By combining VDM (Variational Mode Decomposition), CNN (Convolutional Neural Network), and LSTM (Long Short-Term Memory Network), the present invention can extract multi-frequency features from complex battery operation data and capture the dynamic changes of time series, significantly improving the accuracy of SOC prediction.

[0041] 2. VDM decomposition extracts the signal frequency features, compensating for the defect that traditional neural networks ignore multi-scale characteristics. By combining CNN and LSTM, it can not only extract spatial features but also capture time-dependent relationships. The complete normalization, denoising, and inverse normalization processes ensure data quality and result interpretability.

[0042] 3. The present invention integrates signal processing, deep learning, and time series modeling techniques to construct a complete SOC prediction framework. Compared with traditional Coulomb integration methods or model methods, it reduces the dependence on battery physical models and initial conditions.

[0043] 4. The present invention effectively separates the noise and abnormal components in the signal through modal decomposition, combines the feature extraction and modeling capabilities of the deep learning model, adapts to the SOC prediction requirements under complex working conditions, and has stronger robustness and stability. Description of the Drawings

[0044] Figure 1 It is the estimated curve before VDM processing of the robot lithium battery SOC prediction method based on the VDM-CNN-LSTM neural network according to an embodiment of the present invention;

[0045] Figure 2 It is the estimated curve after VDM processing of the robot lithium battery SOC prediction method based on the VDM-CNN-LSTM neural network according to an embodiment of the present invention;

[0046] Figure 3 It is the overall prediction result of the SOC of the VDM-CNN-LSTM neural network model;

[0047] Figure 4 It is a partial comparison diagram between the VDM-CNN-LSTM neural network model and the LSTM algorithm before improvement;

[0048] Figure 5 It is the prediction error curve of the VDM-CNN-LSTM neural network model;

[0049] Figure 6 It is the flow chart of the present invention. Detailed Embodiments

[0050] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0051] The following illustrates the embodiments of the present invention through specific examples. Those skilled in the art can easily understand the other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0052] The method for predicting the state of charge (SOC) of a robot lithium battery based on a VDM-CNN-LSTM neural network is as follows:

[0053] Step 1, data collection and preprocessing

[0054] Collect the voltage, current, and temperature data of the lithium battery in real time. To ensure the stability of model training and the scale consistency between different features, the time series data needs to be intercepted through a dynamic sliding window and segmented normalization is performed according to the charge and discharge state and the battery temperature. The normalization formula is as shown in Equation (1):

[0055]

[0056] where phase ∈ {charge, discharge, idle}, μ phase (T) and σ phase (T) are the mean and standard deviation of the corresponding working conditions at temperature T, which are dynamically updated through the state of health (SOH) of the battery.

[0057] After obtaining the output result, an inverse normalization process needs to be performed to make the output value meaningful.

[0058] Use the VDM algorithm to decompose signals such as the voltage and current of the lithium battery. The VDM algorithm will attempt to decompose the signal into a series of intrinsic mode functions (IMFs), and each IMF represents the components of different frequencies in the signal. By optimizing a set of parameters to minimize the spectral overlap between adjacent IMFs, the best decomposition effect can be achieved.

[0059] Step 2, adaptively decompose the preprocessed signal through variational mode decomposition (VDM) to obtain multiple intrinsic mode functions (IMFs), and then optimize the parameters through particle swarm optimization (PSO).

[0060] The preprocessed signal is adaptively decomposed using Variational Mode Decomposition (VDM). The number of modes K and the penalty factor α are dynamically adjusted by Particle Swarm Optimization (PSO). The optimization objective is to minimize spectral overlap and maximize the mutual information between the IMF components and the SOC labels. The decomposition objective function is shown in Equation (2):

[0061]

[0062] Among them, the number of modes K is adaptively determined according to the State of Health (SOH) of the battery, as shown in Equation (3):

[0063]

[0064] And, the PSO optimization objective function is shown in Equation (4):

[0065]

[0066] Among them, MI is the mutual information, and λ is the sparsity weight coefficient.

[0067] In PSO optimization, the particle swarm size is set to 50, and the number of iterations is 100 times. The fitness function is as described in Claim 3. The experiment uses the NASA PCoE dataset. The battery charge and discharge cycles cover the environment from -10°C to 45°C. The ratio of the training set to the test set is 8:2. The average RMSE of 5-fold cross-validation is 0.017 ± 0.002.

[0068] Step 3: Take multiple Intrinsic Mode Functions (IMFs) after Variational Mode Decomposition (VDM) as multi-channel inputs. Each IMF is independently input into a Convolutional Neural Network (CNN). Depthwise separable convolutional kernels are used to extract cross-modal coupling features, thereby obtaining the feature sequence output by the CNN;

[0069] Input multiple Intrinsic Mode Functions (IMFs) into a Convolutional Neural Network (CNN). Depthwise separable convolutional kernels are used to extract cross-modal frequency features. The convolution operation formula is shown in Equation (5):

[0070]

[0071] Among them, u m is the m-th IMF, M is the number of selected modes, is the cross-channel convolutional kernel weight.

[0072] And, the number of parameters of the depthwise separable convolutional kernel is shown in Equation (6):

[0073] Params = D k ·M + D k ·D o ·(K ω - 1) (6)

[0074] Among them, D k is the number of input channels, and D o is the number of output channels, and K ω is the size of the convolutional kernel.

[0075] Step 4: Input the feature sequence output by the convolutional neural network (CNN) into the bidirectional LSTM (BiLSTM), and introduce the multi-head self-attention mechanism to strengthen the weights of key time steps;

[0076] Use the bidirectional LSTM (BiLSTM) combined with the multi-head self-attention mechanism to perform time series modeling on the features output by the CNN. The multi-head self-attention mechanism strengthens the weights of key time steps, as shown in Equation (7):

[0077]

[0078] The final hidden state is shown in Equation (8):

[0079]

[0080] Among them, γ is a learnable attention fusion coefficient.

[0081] Although the general trend of the SOC predicted by the LSTM is similar to the true value, there is a large amount of noise in the predicted value. And during the experiment, it is found that there is a randomness problem in the LSTM model fitting the training set, which is not conducive to the optimization process of the particle swarm. Therefore, the VDM algorithm is added on the basis of the LSTM output to decompose the current data, and the effect is as Figure 2 shown. A large amount of noise is filtered out, and the curve also becomes smooth, but the prediction accuracy is still insufficient. This is caused by the selection of LSTM hyperparameters, so convolution is performed.

[0082] Step 5: Map through the fully connected layer and perform inverse normalization to output the SOC prediction value.

[0083] Map to the normalized SOC value through the fully connected layer and perform dynamic inverse normalization based on the real-time temperature, as shown in Equation (9):

[0084]

[0085] Among them, ΔSOC(T) = SOC max (T) - SOC min (T), which is obtained by looking up the table in the battery temperature-capacity curve.

[0086] The local comparison graph with the LSTM algorithm before improvement is as Figure 4 shown. The prediction error curve is as Figure 5As shown, it can be seen that the basic error fluctuates between ±2%, indicating that the improved LSTM model has higher prediction accuracy. From the local magnification Figure 4 and the evaluation indicators in Table 1, it can be seen that the predicted curve of the improved LSTM becomes more stable and smooth. In terms of accuracy, the RMSE after improvement has increased by 56.4% compared with that before improvement, and the MAE has increased by 58.6%. Although the calculation speed has decreased by 15.4%, it still meets the requirements of real-time SOC prediction.

[0087] Table 1 shows the comparison results of establishing a VDM-CNN-LSTM neural network model and a common LSTM model through training. The FUDS working condition is used as the test set and brought into the prediction model of this application to verify the accuracy and stability of the improved model. The overall prediction results of SOC are shown in Table 1.

[0088] Table 1

[0089]

[0090] The above embodiments are only illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes completed by those with ordinary knowledge in the technical field without departing from the spirit and technical idea disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for predicting the state of charge (SOC) of a robot's lithium battery based on a VDM-CNN-LSTM neural network, characterized in that, It includes the following steps: Step 1, data acquisition and preprocessing; Step 2, adaptively decompose the preprocessed signal through variational mode decomposition to obtain multiple intrinsic mode functions, and then optimize the parameters through particle swarm optimization; Step 3, use the multiple intrinsic mode functions after variational mode decomposition as multi-channel inputs. Each IMF is independently input into the convolutional neural network, and depthwise separable convolutional kernels are used to extract cross-modal coupling features, so as to obtain the feature sequence output by the convolutional neural network; Step 4, input the feature sequence output by the convolutional neural network into the bidirectional LSTM, and introduce the multi-head self-attention mechanism to strengthen the weights of key time steps; Step 5, map through the fully connected layer and inverse normalization, and output the SOC prediction value.

Images